libust/lttng-ust-comm.c

   1 /*
   2  * lttng-ust-comm.c
   3  *
   4  * Copyright (C) 2011 David Goulet <david.goulet@polymtl.ca>
   5  * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   6  *
   7  * This library is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; only
  10  * version 2.1 of the License.
  11  *
  12  * This library is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with this library; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 #define _LGPL_SOURCE
  23 #include <sys/types.h>
  24 #include <sys/socket.h>
  25 #include <sys/prctl.h>
  26 #include <sys/mman.h>
  27 #include <sys/stat.h>
  28 #include <sys/types.h>
  29 #include <sys/wait.h>
  30 #include <fcntl.h>
  31 #include <unistd.h>
  32 #include <errno.h>
  33 #include <pthread.h>
  34 #include <semaphore.h>
  35 #include <time.h>
  36 #include <assert.h>
  37 #include <signal.h>
  38 #include <urcu/uatomic.h>
  39 #include <urcu/futex.h>
  40
  41 #include <lttng-ust-comm.h>
  42 #include <ust/usterr-signal-safe.h>
  43 #include <ust/lttng-ust-abi.h>
  44 #include <ust/tracepoint.h>
  45 #include <ust/tracepoint-internal.h>
  46 #include <ust/ust.h>
  47 #include "ltt-tracer-core.h"
  48
  49 /*
  50  * Has lttng ust comm constructor been called ?
  51  */
  52 static int initialized;
  53
  54 /*
  55  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  56  * Held when handling a command, also held by fork() to deal with
  57  * removal of threads, and by exit path.
  58  */
  59
  60 /* Should the ust comm thread quit ? */
  61 static int lttng_ust_comm_should_quit;
  62
  63 /*
  64  * Wait for either of these before continuing to the main
  65  * program:
  66  * - the register_done message from sessiond daemon
  67  *   (will let the sessiond daemon enable sessions before main
  68  *   starts.)
  69  * - sessiond daemon is not reachable.
  70  * - timeout (ensuring applications are resilient to session
  71  *   daemon problems).
  72  */
  73 static sem_t constructor_wait;
  74 /*
  75  * Doing this for both the global and local sessiond.
  76  */
  77 static int sem_count = { 2 };
  78
  79 /*
  80  * Info about socket and associated listener thread.
  81  */
  82 struct sock_info {
  83         const char *name;
  84         pthread_t ust_listener; /* listener thread */
  85         int root_handle;
  86         int constructor_sem_posted;
  87         int allowed;
  88         int global;
  89
  90         char sock_path[PATH_MAX];
  91         int socket;
  92
  93         char wait_shm_path[PATH_MAX];
  94         char *wait_shm_mmap;
  95 };
  96
  97 /* Socket from app (connect) to session daemon (listen) for communication */
  98 struct sock_info global_apps = {
  99         .name = "global",
 100         .global = 1,
 101
 102         .root_handle = -1,
 103         .allowed = 1,
 104
 105         .sock_path = DEFAULT_GLOBAL_APPS_UNIX_SOCK,
 106         .socket = -1,
 107
 108         .wait_shm_path = DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH,
 109 };
 110
 111 /* TODO: allow global_apps_sock_path override */
 112
 113 struct sock_info local_apps = {
 114         .name = "local",
 115         .global = 0,
 116         .root_handle = -1,
 117         .allowed = 0,   /* Check setuid bit first */
 118
 119         .socket = -1,
 120 };
 121
 122 static int wait_poll_fallback;
 123
 124 extern void ltt_ring_buffer_client_overwrite_init(void);
 125 extern void ltt_ring_buffer_client_discard_init(void);
 126 extern void ltt_ring_buffer_metadata_client_init(void);
 127 extern void ltt_ring_buffer_client_overwrite_exit(void);
 128 extern void ltt_ring_buffer_client_discard_exit(void);
 129 extern void ltt_ring_buffer_metadata_client_exit(void);
 130
 131 static
 132 int setup_local_apps(void)
 133 {
 134         const char *home_dir;
 135         uid_t uid;
 136
 137         uid = getuid();
 138         /*
 139          * Disallow per-user tracing for setuid binaries.
 140          */
 141         if (uid != geteuid()) {
 142                 local_apps.allowed = 0;
 143                 return 0;
 144         } else {
 145                 local_apps.allowed = 1;
 146         }
 147         home_dir = (const char *) getenv("HOME");
 148         if (!home_dir)
 149                 return -ENOENT;
 150         snprintf(local_apps.sock_path, PATH_MAX,
 151                  DEFAULT_HOME_APPS_UNIX_SOCK, home_dir);
 152         snprintf(local_apps.wait_shm_path, PATH_MAX,
 153                  DEFAULT_HOME_APPS_WAIT_SHM_PATH, uid);
 154         return 0;
 155 }
 156
 157 static
 158 int register_app_to_sessiond(int socket)
 159 {
 160         ssize_t ret;
 161         int prctl_ret;
 162         struct {
 163                 uint32_t major;
 164                 uint32_t minor;
 165                 pid_t pid;
 166                 pid_t ppid;
 167                 uid_t uid;
 168                 gid_t gid;
 169                 char name[16];  /* process name */
 170         } reg_msg;
 171
 172         reg_msg.major = LTTNG_UST_COMM_VERSION_MAJOR;
 173         reg_msg.minor = LTTNG_UST_COMM_VERSION_MINOR;
 174         reg_msg.pid = getpid();
 175         reg_msg.ppid = getppid();
 176         reg_msg.uid = getuid();
 177         reg_msg.gid = getgid();
 178         prctl_ret = prctl(PR_GET_NAME, (unsigned long) reg_msg.name, 0, 0, 0);
 179         if (prctl_ret) {
 180                 ERR("Error executing prctl");
 181                 return -errno;
 182         }
 183
 184         ret = lttcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 185         if (ret >= 0 && ret != sizeof(reg_msg))
 186                 return -EIO;
 187         return ret;
 188 }
 189
 190 static
 191 int send_reply(int sock, struct lttcomm_ust_reply *lur)
 192 {
 193         ssize_t len;
 194
 195         len = lttcomm_send_unix_sock(sock, lur, sizeof(*lur));
 196         switch (len) {
 197         case sizeof(*lur):
 198                 DBG("message successfully sent");
 199                 return 0;
 200         case -1:
 201                 if (errno == ECONNRESET) {
 202                         printf("remote end closed connection\n");
 203                         return 0;
 204                 }
 205                 return -1;
 206         default:
 207                 printf("incorrect message size: %zd\n", len);
 208                 return -1;
 209         }
 210 }
 211
 212 static
 213 int handle_register_done(struct sock_info *sock_info)
 214 {
 215         int ret;
 216
 217         if (sock_info->constructor_sem_posted)
 218                 return 0;
 219         sock_info->constructor_sem_posted = 1;
 220         if (uatomic_read(&sem_count) <= 0) {
 221                 return 0;
 222         }
 223         ret = uatomic_add_return(&sem_count, -1);
 224         if (ret == 0) {
 225                 ret = sem_post(&constructor_wait);
 226                 assert(!ret);
 227         }
 228         return 0;
 229 }
 230
 231 static
 232 int handle_message(struct sock_info *sock_info,
 233                 int sock, struct lttcomm_ust_msg *lum)
 234 {
 235         int ret = 0;
 236         const struct objd_ops *ops;
 237         struct lttcomm_ust_reply lur;
 238         int shm_fd, wait_fd;
 239
 240         ust_lock();
 241
 242         memset(&lur, 0, sizeof(lur));
 243
 244         if (lttng_ust_comm_should_quit) {
 245                 ret = -EPERM;
 246                 goto end;
 247         }
 248
 249         ops = objd_ops(lum->handle);
 250         if (!ops) {
 251                 ret = -ENOENT;
 252                 goto end;
 253         }
 254
 255         switch (lum->cmd) {
 256         case LTTNG_UST_REGISTER_DONE:
 257                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 258                         ret = handle_register_done(sock_info);
 259                 else
 260                         ret = -EINVAL;
 261                 break;
 262         case LTTNG_UST_RELEASE:
 263                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 264                         ret = -EPERM;
 265                 else
 266                         ret = objd_unref(lum->handle);
 267                 break;
 268         default:
 269                 if (ops->cmd)
 270                         ret = ops->cmd(lum->handle, lum->cmd,
 271                                         (unsigned long) &lum->u);
 272                 else
 273                         ret = -ENOSYS;
 274                 break;
 275         }
 276
 277 end:
 278         lur.handle = lum->handle;
 279         lur.cmd = lum->cmd;
 280         lur.ret_val = ret;
 281         if (ret >= 0) {
 282                 lur.ret_code = LTTCOMM_OK;
 283         } else {
 284                 //lur.ret_code = LTTCOMM_SESSION_FAIL;
 285                 lur.ret_code = ret;
 286         }
 287         switch (lum->cmd) {
 288         case LTTNG_UST_STREAM:
 289                 /*
 290                  * Special-case reply to send stream info.
 291                  * Use lum.u output.
 292                  */
 293                 lur.u.stream.memory_map_size = lum->u.stream.memory_map_size;
 294                 shm_fd = lum->u.stream.shm_fd;
 295                 wait_fd = lum->u.stream.wait_fd;
 296                 break;
 297         case LTTNG_UST_CHANNEL:
 298                 lur.u.channel.memory_map_size = lum->u.channel.memory_map_size;
 299                 shm_fd = lum->u.channel.shm_fd;
 300                 wait_fd = lum->u.channel.wait_fd;
 301                 break;
 302         }
 303         ret = send_reply(sock, &lur);
 304         if (ret < 0) {
 305                 perror("error sending reply");
 306                 goto error;
 307         }
 308
 309         if ((lum->cmd == LTTNG_UST_STREAM || lum->cmd == LTTNG_UST_CHANNEL)
 310                         && lur.ret_code == LTTCOMM_OK) {
 311                 /* we also need to send the file descriptors. */
 312                 ret = lttcomm_send_fds_unix_sock(sock,
 313                         &shm_fd, &shm_fd,
 314                         1, sizeof(int));
 315                 if (ret < 0) {
 316                         perror("send shm_fd");
 317                         goto error;
 318                 }
 319                 ret = lttcomm_send_fds_unix_sock(sock,
 320                         &wait_fd, &wait_fd,
 321                         1, sizeof(int));
 322                 if (ret < 0) {
 323                         perror("send wait_fd");
 324                         goto error;
 325                 }
 326         }
 327 error:
 328         ust_unlock();
 329         return ret;
 330 }
 331
 332 static
 333 void cleanup_sock_info(struct sock_info *sock_info)
 334 {
 335         int ret;
 336
 337         if (sock_info->socket != -1) {
 338                 ret = close(sock_info->socket);
 339                 if (ret) {
 340                         ERR("Error closing apps socket");
 341                 }
 342                 sock_info->socket = -1;
 343         }
 344         if (sock_info->root_handle != -1) {
 345                 ret = objd_unref(sock_info->root_handle);
 346                 if (ret) {
 347                         ERR("Error unref root handle");
 348                 }
 349                 sock_info->root_handle = -1;
 350         }
 351         sock_info->constructor_sem_posted = 0;
 352         if (sock_info->wait_shm_mmap) {
 353                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 354                 if (ret) {
 355                         ERR("Error unmapping wait shm");
 356                 }
 357                 sock_info->wait_shm_mmap = NULL;
 358         }
 359 }
 360
 361 /*
 362  * Using fork to set umask in the child process (not multi-thread safe).
 363  * We deal with the shm_open vs ftruncate race (happening when the
 364  * sessiond owns the shm and does not let everybody modify it, to ensure
 365  * safety against shm_unlink) by simply letting the mmap fail and
 366  * retrying after a few seconds.
 367  * For global shm, everybody has rw access to it until the sessiond
 368  * starts.
 369  */
 370 static
 371 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 372 {
 373         int wait_shm_fd, ret;
 374         pid_t pid;
 375
 376         /*
 377          * Try to open read-only.
 378          */
 379         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 380         if (wait_shm_fd >= 0) {
 381                 goto end;
 382         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 383                 /*
 384                  * Real-only open did not work, and it's not because the
 385                  * entry was not present. It's a failure that prohibits
 386                  * using shm.
 387                  */
 388                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 389                 goto end;
 390         }
 391         /*
 392          * If the open failed because the file did not exist, try
 393          * creating it ourself.
 394          */
 395         pid = fork();
 396         if (pid > 0) {
 397                 int status;
 398
 399                 /*
 400                  * Parent: wait for child to return, in which case the
 401                  * shared memory map will have been created.
 402                  */
 403                 pid = wait(&status);
 404                 if (!WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 405                         wait_shm_fd = -1;
 406                         goto end;
 407                 }
 408                 /*
 409                  * Try to open read-only again after creation.
 410                  */
 411                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 412                 if (wait_shm_fd < 0) {
 413                         /*
 414                          * Real-only open did not work. It's a failure
 415                          * that prohibits using shm.
 416                          */
 417                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 418                         goto end;
 419                 }
 420                 goto end;
 421         } else if (pid == 0) {
 422                 int create_mode;
 423
 424                 /* Child */
 425                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 426                 if (sock_info->global)
 427                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 428                 /*
 429                  * We're alone in a child process, so we can modify the
 430                  * process-wide umask.
 431                  */
 432                 umask(~create_mode);
 433                 /*
 434                  * Try creating shm (or get rw access).
 435                  * We don't do an exclusive open, because we allow other
 436                  * processes to create+ftruncate it concurrently.
 437                  */
 438                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 439                                 O_RDWR | O_CREAT, create_mode);
 440                 if (wait_shm_fd >= 0) {
 441                         ret = ftruncate(wait_shm_fd, mmap_size);
 442                         if (ret) {
 443                                 PERROR("ftruncate");
 444                                 exit(EXIT_FAILURE);
 445                         }
 446                         exit(EXIT_SUCCESS);
 447                 }
 448                 /*
 449                  * For local shm, we need to have rw access to accept
 450                  * opening it: this means the local sessiond will be
 451                  * able to wake us up. For global shm, we open it even
 452                  * if rw access is not granted, because the root.root
 453                  * sessiond will be able to override all rights and wake
 454                  * us up.
 455                  */
 456                 if (!sock_info->global && errno != EACCES) {
 457                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 458                         exit(EXIT_FAILURE);
 459                 }
 460                 /*
 461                  * The shm exists, but we cannot open it RW. Report
 462                  * success.
 463                  */
 464                 exit(EXIT_SUCCESS);
 465         } else {
 466                 return -1;
 467         }
 468 end:
 469         if (wait_shm_fd >= 0 && !sock_info->global) {
 470                 struct stat statbuf;
 471
 472                 /*
 473                  * Ensure that our user is the owner of the shm file for
 474                  * local shm. If we do not own the file, it means our
 475                  * sessiond will not have access to wake us up (there is
 476                  * probably a rogue process trying to fake our
 477                  * sessiond). Fallback to polling method in this case.
 478                  */
 479                 ret = fstat(wait_shm_fd, &statbuf);
 480                 if (ret) {
 481                         PERROR("fstat");
 482                         goto error_close;
 483                 }
 484                 if (statbuf.st_uid != getuid())
 485                         goto error_close;
 486         }
 487         return wait_shm_fd;
 488
 489 error_close:
 490         ret = close(wait_shm_fd);
 491         if (ret) {
 492                 PERROR("Error closing fd");
 493         }
 494         return -1;
 495 }
 496
 497 static
 498 char *get_map_shm(struct sock_info *sock_info)
 499 {
 500         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 501         int wait_shm_fd, ret;
 502         char *wait_shm_mmap;
 503
 504         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 505         if (wait_shm_fd < 0) {
 506                 goto error;
 507         }
 508         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 509                   MAP_SHARED, wait_shm_fd, 0);
 510         /* close shm fd immediately after taking the mmap reference */
 511         ret = close(wait_shm_fd);
 512         if (ret) {
 513                 PERROR("Error closing fd");
 514         }
 515         if (wait_shm_mmap == MAP_FAILED) {
 516                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 517                 goto error;
 518         }
 519         return wait_shm_mmap;
 520
 521 error:
 522         return NULL;
 523 }
 524
 525 static
 526 void wait_for_sessiond(struct sock_info *sock_info)
 527 {
 528         int ret;
 529
 530         ust_lock();
 531         if (lttng_ust_comm_should_quit) {
 532                 goto quit;
 533         }
 534         if (wait_poll_fallback) {
 535                 goto error;
 536         }
 537         if (!sock_info->wait_shm_mmap) {
 538                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 539                 if (!sock_info->wait_shm_mmap)
 540                         goto error;
 541         }
 542         ust_unlock();
 543
 544         DBG("Waiting for %s apps sessiond", sock_info->name);
 545         /* Wait for futex wakeup */
 546         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 547                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 548                         FUTEX_WAIT, 0, NULL, NULL, 0);
 549                 if (ret < 0) {
 550                         if (errno == EFAULT) {
 551                                 wait_poll_fallback = 1;
 552                                 ERR(
 553 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 554 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 555 "Please upgrade your kernel "
 556 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 557 "mainline). LTTng-UST will use polling mode fallback.");
 558                         }
 559                         PERROR("futex");
 560                 }
 561         }
 562         return;
 563
 564 quit:
 565         ust_unlock();
 566         return;
 567
 568 error:
 569         ust_unlock();
 570         return;
 571 }
 572
 573 /*
 574  * This thread does not allocate any resource, except within
 575  * handle_message, within mutex protection. This mutex protects against
 576  * fork and exit.
 577  * The other moment it allocates resources is at socket connexion, which
 578  * is also protected by the mutex.
 579  */
 580 static
 581 void *ust_listener_thread(void *arg)
 582 {
 583         struct sock_info *sock_info = arg;
 584         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 585
 586         /* Restart trying to connect to the session daemon */
 587 restart:
 588         if (prev_connect_failed) {
 589                 /* Wait for sessiond availability with pipe */
 590                 wait_for_sessiond(sock_info);
 591                 if (has_waited) {
 592                         has_waited = 0;
 593                         /*
 594                          * Sleep for 5 seconds before retrying after a
 595                          * sequence of failure / wait / failure. This
 596                          * deals with a killed or broken session daemon.
 597                          */
 598                         sleep(5);
 599                 }
 600                 has_waited = 1;
 601                 prev_connect_failed = 0;
 602         }
 603         ust_lock();
 604
 605         if (lttng_ust_comm_should_quit) {
 606                 ust_unlock();
 607                 goto quit;
 608         }
 609
 610         if (sock_info->socket != -1) {
 611                 ret = close(sock_info->socket);
 612                 if (ret) {
 613                         ERR("Error closing %s apps socket", sock_info->name);
 614                 }
 615                 sock_info->socket = -1;
 616         }
 617
 618         /* Register */
 619         ret = lttcomm_connect_unix_sock(sock_info->sock_path);
 620         if (ret < 0) {
 621                 ERR("Error connecting to %s apps socket", sock_info->name);
 622                 prev_connect_failed = 1;
 623                 /*
 624                  * If we cannot find the sessiond daemon, don't delay
 625                  * constructor execution.
 626                  */
 627                 ret = handle_register_done(sock_info);
 628                 assert(!ret);
 629                 ust_unlock();
 630                 goto restart;
 631         }
 632
 633         sock_info->socket = sock = ret;
 634
 635         /*
 636          * Create only one root handle per listener thread for the whole
 637          * process lifetime.
 638          */
 639         if (sock_info->root_handle == -1) {
 640                 ret = lttng_abi_create_root_handle();
 641                 if (ret) {
 642                         ERR("Error creating root handle");
 643                         ust_unlock();
 644                         goto quit;
 645                 }
 646                 sock_info->root_handle = ret;
 647         }
 648
 649         ret = register_app_to_sessiond(sock);
 650         if (ret < 0) {
 651                 ERR("Error registering to %s apps socket", sock_info->name);
 652                 prev_connect_failed = 1;
 653                 /*
 654                  * If we cannot register to the sessiond daemon, don't
 655                  * delay constructor execution.
 656                  */
 657                 ret = handle_register_done(sock_info);
 658                 assert(!ret);
 659                 ust_unlock();
 660                 goto restart;
 661         }
 662         ust_unlock();
 663
 664         for (;;) {
 665                 ssize_t len;
 666                 struct lttcomm_ust_msg lum;
 667
 668                 len = lttcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 669                 switch (len) {
 670                 case 0: /* orderly shutdown */
 671                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 672                         goto end;
 673                 case sizeof(lum):
 674                         DBG("message received\n");
 675                         ret = handle_message(sock_info, sock, &lum);
 676                         if (ret < 0) {
 677                                 ERR("Error handling message for %s socket", sock_info->name);
 678                         }
 679                         continue;
 680                 case -1:
 681                         if (errno == ECONNRESET) {
 682                                 ERR("%s remote end closed connection\n", sock_info->name);
 683                                 goto end;
 684                         }
 685                         goto end;
 686                 default:
 687                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 688                         continue;
 689                 }
 690
 691         }
 692 end:
 693         goto restart;   /* try to reconnect */
 694 quit:
 695         return NULL;
 696 }
 697
 698 /*
 699  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 700  */
 701 static
 702 int get_timeout(struct timespec *constructor_timeout)
 703 {
 704         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 705         char *str_delay;
 706         int ret;
 707
 708         str_delay = getenv("UST_REGISTER_TIMEOUT");
 709         if (str_delay) {
 710                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 711         }
 712
 713         switch (constructor_delay_ms) {
 714         case -1:/* fall-through */
 715         case 0:
 716                 return constructor_delay_ms;
 717         default:
 718                 break;
 719         }
 720
 721         /*
 722          * If we are unable to find the current time, don't wait.
 723          */
 724         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 725         if (ret) {
 726                 return -1;
 727         }
 728         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 729         constructor_timeout->tv_nsec +=
 730                 (constructor_delay_ms % 1000UL) * 1000000UL;
 731         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 732                 constructor_timeout->tv_sec++;
 733                 constructor_timeout->tv_nsec -= 1000000000UL;
 734         }
 735         return 1;
 736 }
 737
 738 /*
 739  * sessiond monitoring thread: monitor presence of global and per-user
 740  * sessiond by polling the application common named pipe.
 741  */
 742 /* TODO */
 743
 744 void __attribute__((constructor)) lttng_ust_init(void)
 745 {
 746         struct timespec constructor_timeout;
 747         int timeout_mode;
 748         int ret;
 749
 750         if (uatomic_xchg(&initialized, 1) == 1)
 751                 return;
 752
 753         /*
 754          * We want precise control over the order in which we construct
 755          * our sub-libraries vs starting to receive commands from
 756          * sessiond (otherwise leading to errors when trying to create
 757          * sessiond before the init functions are completed).
 758          */
 759         init_usterr();
 760         init_tracepoint();
 761         ltt_ring_buffer_metadata_client_init();
 762         ltt_ring_buffer_client_overwrite_init();
 763         ltt_ring_buffer_client_discard_init();
 764
 765         timeout_mode = get_timeout(&constructor_timeout);
 766
 767         ret = sem_init(&constructor_wait, 0, 0);
 768         assert(!ret);
 769
 770         ret = setup_local_apps();
 771         if (ret) {
 772                 ERR("Error setting up to local apps");
 773         }
 774         ret = pthread_create(&local_apps.ust_listener, NULL,
 775                         ust_listener_thread, &local_apps);
 776
 777         if (local_apps.allowed) {
 778                 ret = pthread_create(&global_apps.ust_listener, NULL,
 779                                 ust_listener_thread, &global_apps);
 780         } else {
 781                 handle_register_done(&local_apps);
 782         }
 783
 784         switch (timeout_mode) {
 785         case 1: /* timeout wait */
 786                 do {
 787                         ret = sem_timedwait(&constructor_wait,
 788                                         &constructor_timeout);
 789                 } while (ret < 0 && errno == EINTR);
 790                 if (ret < 0 && errno == ETIMEDOUT) {
 791                         ERR("Timed out waiting for ltt-sessiond");
 792                 } else {
 793                         assert(!ret);
 794                 }
 795                 break;
 796         case -1:/* wait forever */
 797                 do {
 798                         ret = sem_wait(&constructor_wait);
 799                 } while (ret < 0 && errno == EINTR);
 800                 assert(!ret);
 801                 break;
 802         case 0: /* no timeout */
 803                 break;
 804         }
 805 }
 806
 807 static
 808 void lttng_ust_cleanup(int exiting)
 809 {
 810         cleanup_sock_info(&global_apps);
 811         if (local_apps.allowed) {
 812                 cleanup_sock_info(&local_apps);
 813         }
 814         lttng_ust_abi_exit();
 815         ltt_events_exit();
 816         ltt_ring_buffer_client_discard_exit();
 817         ltt_ring_buffer_client_overwrite_exit();
 818         ltt_ring_buffer_metadata_client_exit();
 819         exit_tracepoint();
 820         if (!exiting) {
 821                 /* Reinitialize values for fork */
 822                 sem_count = 2;
 823                 lttng_ust_comm_should_quit = 0;
 824                 initialized = 0;
 825         }
 826 }
 827
 828 void __attribute__((destructor)) lttng_ust_exit(void)
 829 {
 830         int ret;
 831
 832         /*
 833          * Using pthread_cancel here because:
 834          * A) we don't want to hang application teardown.
 835          * B) the thread is not allocating any resource.
 836          */
 837
 838         /*
 839          * Require the communication thread to quit. Synchronize with
 840          * mutexes to ensure it is not in a mutex critical section when
 841          * pthread_cancel is later called.
 842          */
 843         ust_lock();
 844         lttng_ust_comm_should_quit = 1;
 845         ust_unlock();
 846
 847         ret = pthread_cancel(global_apps.ust_listener);
 848         if (ret) {
 849                 ERR("Error cancelling global ust listener thread");
 850         }
 851         if (local_apps.allowed) {
 852                 ret = pthread_cancel(local_apps.ust_listener);
 853                 if (ret) {
 854                         ERR("Error cancelling local ust listener thread");
 855                 }
 856         }
 857         lttng_ust_cleanup(1);
 858 }
 859
 860 /*
 861  * We exclude the worker threads across fork and clone (except
 862  * CLONE_VM), because these system calls only keep the forking thread
 863  * running in the child.  Therefore, we don't want to call fork or clone
 864  * in the middle of an tracepoint or ust tracing state modification.
 865  * Holding this mutex protects these structures across fork and clone.
 866  */
 867 void ust_before_fork(ust_fork_info_t *fork_info)
 868 {
 869         /*
 870          * Disable signals. This is to avoid that the child intervenes
 871          * before it is properly setup for tracing. It is safer to
 872          * disable all signals, because then we know we are not breaking
 873          * anything by restoring the original mask.
 874          */
 875         sigset_t all_sigs;
 876         int ret;
 877
 878         /* Disable signals */
 879         sigfillset(&all_sigs);
 880         ret = sigprocmask(SIG_BLOCK, &all_sigs, &fork_info->orig_sigs);
 881         if (ret == -1) {
 882                 PERROR("sigprocmask");
 883         }
 884         ust_lock();
 885         rcu_bp_before_fork();
 886 }
 887
 888 static void ust_after_fork_common(ust_fork_info_t *fork_info)
 889 {
 890         int ret;
 891
 892         DBG("process %d", getpid());
 893         ust_unlock();
 894         /* Restore signals */
 895         ret = sigprocmask(SIG_SETMASK, &fork_info->orig_sigs, NULL);
 896         if (ret == -1) {
 897                 PERROR("sigprocmask");
 898         }
 899 }
 900
 901 void ust_after_fork_parent(ust_fork_info_t *fork_info)
 902 {
 903         DBG("process %d", getpid());
 904         rcu_bp_after_fork_parent();
 905         /* Release mutexes and reenable signals */
 906         ust_after_fork_common(fork_info);
 907 }
 908
 909 /*
 910  * After fork, in the child, we need to cleanup all the leftover state,
 911  * except the worker thread which already magically disappeared thanks
 912  * to the weird Linux fork semantics. After tyding up, we call
 913  * lttng_ust_init() again to start over as a new PID.
 914  *
 915  * This is meant for forks() that have tracing in the child between the
 916  * fork and following exec call (if there is any).
 917  */
 918 void ust_after_fork_child(ust_fork_info_t *fork_info)
 919 {
 920         DBG("process %d", getpid());
 921         /* Release urcu mutexes */
 922         rcu_bp_after_fork_child();
 923         lttng_ust_cleanup(0);
 924         /* Release mutexes and reenable signals */
 925         ust_after_fork_common(fork_info);
 926         lttng_ust_init();
 927 }